Datasets:
π logical-fallacy-classification v1.0 β 183 hand-authored examples, 20 fallacy types
Dataset Published: logical-fallacy-classification v1.0
I've just published logical-fallacy-classification, a curated benchmark for detecting and classifying logical fallacies in natural language text.
What's in it
| Property | Details |
|---|---|
| Total examples | 183 |
| Fallacy types | 20 canonical informal logic fallacies |
| Negative class | 20 sound (non-fallacious) examples |
| Domains | politics, science, health, law, economics, ethics, philosophy, education, marketing, everyday |
| Difficulty levels | subtle / moderate / overt |
| Splits | train (137) / validation (23) / test (23) |
| Formats | JSONL + Parquet |
| License | Apache 2.0 |
Fallacy types covered
ad_hominem, straw_man, slippery_slope, false_dilemma, appeal_to_authority, hasty_generalization, post_hoc, circular_reasoning, appeal_to_emotion, appeal_to_popularity, false_equivalence, red_herring, appeal_to_nature, equivocation, sunk_cost, anecdotal_evidence, no_true_scotsman, tu_quoque, appeal_to_ignorance, naturalistic_fallacy
Quick start
from datasets import load_dataset
ds = load_dataset("SalutefaceAAC/logical-fallacy-classification")
# Binary: fallacious vs. sound
for ex in ds["train"]:
label = 1 if ex["label"] == "fallacious" else 0
# 20-class multi-label classification
types = set(ex["fallacy_type"] for ex in ds["train"])
# Filter to subtle examples (require domain knowledge to spot)
subtle = ds["train"].filter(lambda x: x["difficulty"] == "subtle")
Why this dataset
Most LLM benchmarks test factual recall or instruction following. Identifying why an argument is logically flawed requires a different kind of reasoning β understanding the argument's structure, not just its surface content. This dataset is intended to:
- Fine-tune classifiers for fallacy detection in political speech, health misinformation, and marketing copy
- Evaluate whether LLMs can correctly label and explain logical errors
- Serve as a component in fact-checking and misinformation detection pipelines
- Support educational tools for rhetoric and argumentation courses
Design notes
All examples are hand-authored rather than scraped or LLM-generated in bulk. Each includes:
- A verbose
explanationfield (why it is/isn't a fallacy β useful for explanation generation tasks) - A
subcategoryfield for finer-grained variants (e.g.,tu_quoquewithinad_hominem) - Balanced difficulty across the three levels
Feedback, issues, and PRs welcome β especially additional examples for underrepresented categories (marketing, philosophy).
Published by SalutefaceAAC as part of the SalutefaceAAC OSS dataset portfolio.
Usage tip: separating detection from classification tasks
A few researchers have reached out about how to use this dataset for fine-tuning. One practical note:
Detection vs. classification are different tasks and this dataset supports both, but the framing matters:
For detection (binary: fallacy or not):
from datasets import load_dataset
ds = load_dataset("SalutefaceAAC/logical-fallacy-classification")
# The 20 "sound" examples (fallacy_type == "sound") are your negatives
# Everything else is positive
ds = ds.map(lambda x: {"label": 0 if x["fallacy_type"] == "sound" else 1})
For 20-class classification:
Fine-tune only on the positive examples (~163 labeled by type). If your application sees many more negatives at inference time, augment training negatives proportionally.
On evaluation
Two common mistakes I've seen with fallacy detection datasets:
- Testing only on text that clearly contains fallacies β this inflates precision dramatically. Always test on a held-out set that includes sound arguments.
- Lumping rare types together β ad hominem has 10 examples; composition fallacy has 6. Report per-type metrics rather than macro averages, or aggregate rare types into an "other" bin.
The severity field (1β5) is useful for ranking outputs in an LLM pipeline β filter to high-severity positives first for human review.
What's planned
Looking at expanding to ~300 examples with emphasis on:
- More informal fallacies that require real-world knowledge (no true Scotsman, special pleading)
- Multi-step arguments where the fallacy isn't in a single sentence
- Examples in non-English contexts (though the current v1.0 is English-only)
Happy to discuss applications or evaluation methodology.